How to Access GPT and Claude in Hong Kong: Every Legitimate Route (2026)

A Hong Kong business can legitimately use OpenAI’s GPT-5.6 and newer GPT-6 models and Anthropic’s Claude Opus 5 today, without a VPN and without breaching anyone’s terms: GPT through Microsoft 365 Copilot and Microsoft Foundry, and Claude through Copilot’s model picker (where your tenant offers it) and Perplexity. That surprises a lot of people, because the headline experience hasn’t changed: ChatGPT.com still errors from a Hong Kong IP, and claude.ai still turns you away. Both are vendor geo-blocks, not government bans (our guide to whether ChatGPT is banned in Hong Kong covers that distinction). This post is the companion piece: not whether the models are blocked, but how to reach them anyway, through the front door. As always with this topic, treat it as a snapshot (September 2026): availability moves fast, and we update these guides when it does. If what you want is these models inside your own firm’s Microsoft 365, our Copilot rollout service is the practical next step.

Is Claude available in Hong Kong? The short answer

Not first-party: Anthropic does not support Hong Kong, so claude.ai and the Claude API refuse Hong Kong users, and AWS Bedrock and Microsoft Foundry both refuse Hong Kong billing accounts for Claude. The governed route is Claude inside Microsoft 365 Copilot where your tenant offers it, which is the route the rest of this guide walks through.

Every legitimate route at a glance

Route	What you get	Frontier GPT?	Frontier Claude?	Best for
Microsoft 365 Copilot	AI inside Word, Excel, Teams, Outlook	✅ GPT-5.6 family, GPT-6 rolling out	✅ Claude Opus 5, Opus 5.5 rolling out (tenant dependent)	Most Hong Kong businesses
Microsoft Foundry (Azure)	The models as APIs, for your own apps	✅ GPT-5.6 and GPT-6 families	❌ Needs a billing account in a country Anthropic supports	Developers, custom apps
Amazon Bedrock (AWS)	Model APIs on AWS	❌ OpenAI’s Bedrock terms exclude HK (open-weight gpt-oss only)	❌ Blocked for HK accounts and connections	AWS shops using open-weight and Amazon models
Google Vertex AI / Gemini	Gemini direct; partner models via API	➖ No GPT	⚠️ Claude listed, HK access unverified	Google Cloud shops; Gemini users
Perplexity	Research assistant with model choice	✅ GPT-5.6 (Sol on Max)	✅ Sonnet 5 on Pro, Opus 5 on Max	Individuals, research work
Poe and other aggregators	Multi-model chat apps	✅	✅	Casual use; check terms first
Cloudflare AI Gateway	Proxy, billing and caching layer	➖ Depends on upstream provider	➖ Same caveat	Developers already on Cloudflare

The rest of this guide walks through each route, what it costs, and the caveats that matter for a business.

How to use GPT-5.6 in Hong Kong: Microsoft 365 Copilot

The simplest legitimate route to OpenAI’s frontier models in Hong Kong is the one most businesses already half-own. Microsoft 365 Copilot is fully supported here, and since July 2026 it has run OpenAI’s GPT-5.6 family: three tiers named Sol (the most capable), Terra (balanced) and Luna (fast). Copilot doesn’t show you those names; they sit behind its response-mode labels. OpenAI’s next generation is arriving the same way. GPT-6 Astra, OpenAI’s most capable model, has been selectable in Copilot Cowork and Copilot Studio since 4 September, and GPT-6 Sol began rolling out across Word, Excel, PowerPoint and Chat on 22 September. Microsoft says availability may vary by licence and region, but the ceiling is the real thing: the models OpenAI sells in supported countries, running inside Word, Excel, PowerPoint, Outlook and Teams, on a Hong Kong tenant, with no workaround.

Because Copilot is licensed through Microsoft rather than OpenAI, none of OpenAI’s regional exclusions apply. You also get what the consumer app can’t give you: enterprise data protection (prompts and responses are not used to train the models), plus DLP, sensitivity labels and audit through standard Microsoft 365 admin tooling. Whether the licence is worth it per user is a separate question, and we’ve written a full assessment of whether Copilot is worth it. Rolled out properly (see Microsoft 365 setup), this is the default answer for most Hong Kong organisations.

How to use Claude in Hong Kong: narrower, but real

Anthropic is the stricter of the two vendors. Hong Kong isn’t on its supported-countries list, the restriction covers the API as well as claude.ai, and enforcement tightened through 2026: Goldman Sachs and JPMorgan are among the firms that have withdrawn Claude from their Hong Kong staff, and Anthropic has said Claude was never officially supported here. Its terms also reach past geography: since September 2025 Anthropic has refused companies more than 50% owned by groups headquartered in regions it doesn’t support, Hong Kong and the Mainland included, wherever they operate, so a Hong Kong group can’t simply sign up through its Singapore subsidiary. So the widely-shared assumption (“you just can’t use Claude here”) is closer to true than it is for GPT. Closer, but not correct. Two legitimate doors are open:

Microsoft 365 Copilot’s model picker. Microsoft added Anthropic models to Copilot in late 2025 and has kept pace since: Claude Opus 5 has been rolling out in Copilot since late July 2026, and Claude Opus 5.5 followed on 22 September, selectable in Word, Excel, PowerPoint, Copilot Chat and Copilot Cowork, Microsoft’s agentic co-working mode where the model works through a multi-step task alongside you rather than answering one prompt at a time. Two caveats belong in any honest write-up. First, Microsoft’s documentation switches Anthropic models on by default for most commercial tenants and names only the EU, EFTA, UK and its government and sovereign clouds as exceptions. It says nothing about Hong Kong either way, and Anthropic itself doesn’t list Hong Kong, so check your own tenant’s model picker before you build a plan on it. Second, Claude in Copilot runs with Anthropic as a Microsoft subprocessor: Microsoft’s Product Terms and Data Protection Addendum apply, but processing falls outside Microsoft’s in-country and EU Data Boundary commitments, and some advanced models (Claude Fable 5, for one) stay off until an admin accepts Anthropic’s own terms. That’s a governance consideration your compliance team should sign off on.

Perplexity. Claude Sonnet 5 is selectable on Perplexity Pro, and Claude Opus 5 on Perplexity Max. More on pricing below.

What does not work from Hong Kong, despite what you may have read: Anthropic’s own apps and API, including the Claude Cowork desktop app, and (more surprisingly) Claude on AWS Bedrock and Microsoft Foundry, covered below.

Amazon Bedrock in Hong Kong: what works and what doesn’t

Amazon Bedrock is AWS’s managed service for foundation models, and on paper it looks like the obvious Claude route: Claude Opus 5 launched on AWS in July 2026. In practice, two things limit it from Hong Kong:

• No Anthropic models in the Hong Kong region (ap-east-1). The nearest Claude endpoints are in Singapore, Tokyo, Seoul and Sydney. That part is routine; plenty of HK workloads already run in Singapore.
• Anthropic’s geo-restriction passes through Bedrock. AWS’s own guidance on Claude access (reviewed 15 September 2026) says calls from countries, regions or territories Anthropic doesn’t support fail with an “Access to Anthropic models is not allowed” error, and that the AWS account and billing address must also be in a supported country. Bedrock checks where the call comes from as well as where the account is registered, so a Hong Kong business is refused whichever AWS region serves the model.

Nor is Bedrock a way to reach OpenAI’s frontier models from Hong Kong, although they are on it now: GPT-5.4 and GPT-5.5 since June 2026, the GPT-5.6 family since July, and GPT-6 Astra, Sol and Luna since September. OpenAI sells them there itself, under its own agreement that you accept by calling a model, and that agreement bars customers and their users from accessing the models outside OpenAI’s supported countries. Hong Kong isn’t on the list. An AWS support engineer’s troubleshooting note tells anyone who hits an unsupported-country error on these models to check the account’s tax and billing addresses and where the call comes from, and warns that switching regions won’t get round it. There’s no Asian data residency either: the models can be kept to a single region only in the US, and from Singapore or Tokyo the GPT-6 and GPT-5.6 models are reachable only through AWS’s global routing, which can process a request in any commercial region. For GPT from Hong Kong, use Copilot or Foundry.

Bedrock is still far from useless here. OpenAI’s open-weight gpt-oss models (on Bedrock since August 2025) are sold by AWS under their open licence rather than OpenAI’s agreement, and they sit alongside Amazon’s own Nova family, Meta’s Llama, Mistral and others. For an AWS-first Hong Kong business, Bedrock remains a strong platform for those models; just don’t pick it for Claude or frontier GPT.

Microsoft Foundry: the developer route to GPT (not Claude)

If you want the models as APIs rather than inside Office, Microsoft Foundry (the platform formerly known as Azure AI Foundry / Azure OpenAI Service) is the developer-grade equivalent of the Copilot story. Hong Kong isn’t an Azure OpenAI deployment region, but that’s a data-centre detail, not an access block: a Hong Kong organisation can subscribe and deploy the models in a nearby region such as Singapore, Japan or Korea, and call them from Hong Kong. The catalogue includes the full GPT-5.6 family and, since September, GPT-6 Astra, Sol and Luna, which means a HK developer can build directly against the same flagship models OpenAI won’t sell here first-party.

Anthropic’s Claude models also went generally available in Microsoft Foundry in June 2026, in both Anthropic-hosted and Azure-hosted flavours, but they aren’t a Hong Kong route. Microsoft’s deployment guide requires a billing account in a country where Anthropic sells the models and warns that Anthropic’s Supported Regions Policy may apply, so the restriction passes through just as it does on Bedrock. Treat Foundry as the GPT route it is, not a Claude route.

Google: Gemini works directly, Vertex AI for the API

Google took the opposite path to OpenAI and Anthropic: Gemini has been officially available to Hong Kong users since March 2026, no workaround required, which makes it the only frontier US assistant you can simply sign up for here. For developers, Vertex AI exposes Gemini via API from Google Cloud regions near Hong Kong, and its Model Garden lists partner models including Claude Opus 5 and 5.5. Google’s documentation doesn’t say whether Hong Kong customers can use them, and we haven’t verified Claude on Vertex from a Hong Kong connection; with AWS and Microsoft both enforcing Anthropic’s policy, don’t plan on it. Gemini itself carries no such asterisk.

Perplexity, Poe and the aggregator route

Perplexity works directly in Hong Kong (HKT even bundled twelve months of Pro for its mobile and broadband customers during 2025–26, though that enrolment window has closed) and remains the easiest way for an individual to choose which frontier model answers them. The catch is the paywall geometry. The US$20/month Pro tier includes GPT-5.6 Terra and Claude Sonnet 5, which is genuine frontier access at a reasonable price. But the flagship tiers, GPT-5.6 Sol and Claude Opus 5, sit on Perplexity Max at US$200/month per person. For one power user, fine; as a team’s primary AI tool, that’s Copilot money without Copilot’s tenant governance, so most businesses should treat Perplexity as a complement rather than the platform.

Poe (Quora’s multi-model app) is accessible in Hong Kong and offers GPT and Claude models under one subscription. OpenRouter does the same for developers as an OpenAI-compatible API over hundreds of models. Both are legitimate businesses, but they sit further from the model vendors than the Microsoft and AWS routes do, so check the current terms (and where your data goes) before putting work through them. Our comparison of which AI tools work in Hong Kong and China goes deeper on the data-governance side.

Cloudflare’s AI platform deserves a mention because it’s often suggested as an access route, and that’s half right. Workers AI hosts open-weight models (some of them frontier-class, none of them GPT or Claude), and AI Gateway, unified with Workers AI in August 2026, lets developers route, cache and even pre-pay for calls to OpenAI, Anthropic and others through one control plane. It’s a genuinely useful developer layer. But it’s a proxy, not a licence: whether an upstream provider serves a Hong Kong-origin call is still that provider’s decision, and Cloudflare doesn’t change it.

Why not just use a VPN?

Because every route above exists precisely so you don’t have to. A personal ChatGPT or Claude account run through a VPN breaches the provider’s terms, can be cut off without notice, and puts whatever your staff paste into it entirely outside your governance. Hong Kong’s regulators don’t ban AI tools, but the PCPD’s guidance expects organisations to control what data enters them, and for licensed firms the SFC expects the same in regulated work. Shadow AI over a VPN fails all of that at once. The fix is the one that runs through this whole guide: give people a sanctioned route that’s actually good, then back it with policy and security controls.

What a Hong Kong business should actually do

1. Already on Microsoft 365? Copilot is the answer for most: frontier GPT by default, Claude Opus 5 or 5.5 in the model picker if your tenant shows it, enterprise data protection throughout. Confirm the Claude option on your tenant before promising it to anyone.
2. Building products? Microsoft Foundry from a nearby Azure region for GPT; Vertex for Gemini; Bedrock or Vertex for open-weight models. Don’t architect around Claude from Hong Kong: Anthropic’s API, Bedrock and Foundry all refuse Hong Kong accounts. The same goes for GPT on Bedrock, where OpenAI’s terms exclude Hong Kong.
3. Individuals and researchers: Perplexity Pro is the best US$20 in Hong Kong AI; step up to Max only if you specifically need Sol or Opus 5.
4. Whatever the route, govern it: an AI acceptable-use policy, DLP behind it, and no personal, confidential or regulated data into consumer tiers. That, more than model choice, is what separates safe adoption from an incident report.

PTS helps Hong Kong businesses adopt AI safely: route and licensing decisions, Copilot rollout, acceptable-use policy, DLP and data classification, and the cross-border considerations for firms with Mainland operations. If you want a practical, costed plan rather than a model shoot-out, talk to us.

FAQ: accessing frontier AI models from Hong Kong

Can I use GPT-5.6 in Hong Kong?

Yes. OpenAI still doesn’t serve Hong Kong first-party, but the GPT-5.6 family (Sol, Terra and Luna) is fully available here through Microsoft: inside Microsoft 365 Copilot for everyday work, and as APIs through Microsoft Foundry deployed in a nearby Azure region. The newer GPT-6 models are arriving through the same two channels. Perplexity also offers GPT-5.6 (Terra on Pro, Sol on Max). Amazon Bedrock lists the same models, but OpenAI’s terms there limit them to its supported countries, so it isn’t a Hong Kong route.

Is Claude available in Hong Kong?

Not from Anthropic directly: claude.ai and the Claude API both refuse Hong Kong. But yes, through two doors: Microsoft 365 Copilot’s model picker (including Copilot Cowork, with Opus 5.5 now rolling out), subject to tenant and region settings you should verify on your own tenant, and Perplexity Max. Anthropic’s own apps and API remain unavailable here, and the restriction follows Claude onto the cloud platforms: AWS says Bedrock refuses Hong Kong accounts and connections, and Microsoft Foundry sells Claude only to billing accounts in countries Anthropic supports.

Is it legal to access these models from Hong Kong this way?

Yes. These aren’t loopholes; they’re licensed distribution channels. Microsoft, Google, Perplexity and Poe all serve Hong Kong customers under their own terms, and the model vendors are paid through those channels. The catch is that a vendor’s own terms can travel with its model, as Anthropic’s do on Bedrock and Foundry and OpenAI’s do on Bedrock, so those combinations aren’t Hong Kong routes. What breaches terms is going around the geo-block on the vendor’s first-party service with a VPN.

Why do OpenAI and Anthropic block Hong Kong at all?

Neither company has published detailed reasoning for its regional restrictions; both simply exclude Hong Kong from their supported-countries lists, decisions generally read as regulatory and geopolitical caution rather than anything Hong Kong has done. The practical point for a business is that it’s a vendor decision, not a Hong Kong ban: the territory’s internet is open, and both vendors’ models still reach Hong Kong through partners such as Microsoft.

Related reading: Is ChatGPT banned in Hong Kong? · AI tools compared: what works in Hong Kong & China · Is Microsoft 365 Copilot worth it? · Practical AI advisory for Hong Kong & China · Microsoft 365 in China · AI scams in Hong Kong